| Objective: Deep learning network is applied to classification models in various fields,which requires a large number of sample data as support to train the network,so as to obtain good results.When deep learning algorithm is used to classify ECG data,the heart is a very complex organ in the human body,so there are many kinds of sample data of cardiovascular diseases,but the number of samples of many rare diseases is small,and the collection process is difficult.In the training,the deep learning network will overfit due to the small number of samples.The purpose of this study is to use the method of small sample element learning,through a small number of samples,to study the classification network of ECG data with a small number of samples,and put forward a ECG data classification model based on the twin network framework,so that the model can achieve good classification effect under the condition of a small number of ECG data samples.Methods: Based on the framework of twin networks,this study designed a SCLA twin network model with convolutional neural networks and cyclic neural networks and attention mechanisms as subnetworks.The twin network model is composed of two network branches with the same weight and structure.We design a single network branch into three parts and splice the three parts in series as a single network branch in the twin network model,that is,the subnetwork in the twin network model.The first part is the convolutional neural network composed of three convolutional modules,the second part is the LSTM part composed of the special cyclic neural network Bi LSTM,and the third part is the attention mechanism part.Three ablation experimental models were designed,namely,SCNN,a twin network model with convolutional neural network branch,SLSTM,a twin network model with LSTM branch,and SCLSTM,a twin network model with convolutional neural network and LSTM branch.In the comparison method,the SCLA model designed in this study was compared with the matching network model and the prototype network model.Results: We used the data from the PTB-XL data set to train and test the SCLA model of the twin network,and designed three ablation experimental models to train and test under the same data set.SCLA model achieved good results,the accuracy rate was 0.7133,the accuracy rate was 0.7183,the recall rate was 0.7143,and the F1 score was 0.7163 in the PTB-XL data set.The accuracy rate of the three comparison models in the ablation experiment was 0.6633,0.6767,0.5430;the accuracy rate was 0.6673,0.6892,0.5472;the recall rate was 0.6650,0.6785,0.5469;and the F1 score was 0.6661,0.6838,0.5470.SCLA model has better classification effect in ablation experiment.Then,the SCLA model was compared with the matched network model and prototype network model in the small sample learning method based on meta-learning.The SCLA model was superior to the matched network model and prototype network model in four evaluation indexes.Finally,the three groups of models are trained under different sample numbers.The results show that the classification accuracy of the three groups of models increases with the increase of sample size,and the best classification effect is achieved when the number of samples is 15.Conclusions: Based on the results of ablation and comparison experiments,the SCLA twin network model designed in this study achieved relatively high accuracy in classifying ECG data.This not only shows that the combination of convolutional neural network with LSTM and attention mechanism can highlight its own feature extraction advantages,and play a positive role in the classification of small samples of ECG data by twin networks.At the same time,it is verified that the performance of the twin network model is better than that of prototype network and matching network,which are small sample learning methods based on meta-learning,when dealing with the classification of small sample ECG data.It has a good application prospect in the intelligent detection of some rare cardiovascular diseases. |
PDF Full Text Request